Adversarial Audio and Hidden Commands
語音對抗與隱藏指令
Verification & Validation
Risk Description
When hidden commands are embedded in audio played in a public space; due to unmitigated control gaps, nearby devices with voice interfaces receive and execute unauthorized operations, triggering external stakeholder impacts and causing users perceive nothing unusual, and the source of the commands is difficult to trace afterward.
Framework Mappings
MITRE ATLASAML.T0043
NIST AI 100-2Evasion
OWASP Top 10 for LLMLLM01
ISO/IEC 5338驗證與確效
MIT AI Risk RepositoryDomain 2
Risk Treatment & Implementation Guidance
Deploy anomalous-signal detection on voice interfaces to identify hidden commands outside human-audible ranges; Require multi-factor confirmation for sensitive operations so voice alone cannot trigger high-risk actions; Retain command provenance and audio signatures for after-the-fact tracing